Dynamically Grown Generative Adversarial Networks

Recent work introduced progressive network growing as a promising way to ease the training for large GANs, but model design and architecture-growing strategy still remain under-explored needs manual different image data. In this paper, we propose method dynamically grow GAN during training, optimizing architecture its parameters together with automation. The embeds search techniques an interleaving step gradient-based periodically seek optimal generator discriminator. It enjoys benefits of both eased because improved performance broader space. Experimental results demonstrate new state-of-the-art generation. Observations in procedure also provide constructive insights into such generator-discriminator balance convolutional layer choices.